GitHub's Trending Page Is Now a Catalog of Agent Infrastructure

Open GitHub trending on a random day this week and you will not see the usual mix of frameworks and hobby projects at the top. You will see plumbing. The repositories climbing fastest are the parts that sit underneath AI agents: memory that survives a session, teams of agents that share state, marketplaces for their skills, and scanners that check those skills for trouble before they run.

That is a shift in what open-source contributors consider urgent. A year ago, the trending page rewarded new apps. Now it rewards the infrastructure that apps are built on, and the fastest stars are going to the layer most people never see.

The reason is that the agent quietly became the default way to talk to software. Once people started running coding agents, research agents and work agents as a matter of routine, the missing pieces stopped being better models and started being everything around them. A model that forgets, cannot coordinate with another model, or runs code it should not is not a model problem. It is an infrastructure problem, and the trending page is where the infrastructure crowd shows up to fix it.

Memory is the loudest cluster

The clearest signal is persistent memory. thedotmack/claude-mem, a plugin that captures what an agent does during a session, compresses it, and injects the relevant parts back into later sessions, sits at roughly 99,000 stars and gained more than 4,000 in a single week. It works across Claude Code, Codex, Gemini, Copilot and OpenCode, which tells you the problem it solves is not specific to one vendor. Agents forget, and users are tired of re-explaining context every time they open a new window.

Around that flagship sit several tools attacking the same problem from different angles. mksglu/context-mode claims a 98% reduction in tool-output tokens by sandboxing and routing context through the Model Context Protocol. ranxianglei/billion-context promises five times fewer tokens and month-long sessions by compressing context rather than truncating it. modelcontextprotocol/servers provides the base layer that much of this relies on.

The reason memory leads is simple. An agent that cannot remember yesterday cannot be trusted with anything that spans more than one sitting. Every serious use of an agent eventually needs continuity, and continuity was the missing piece for most of the past year.

Orchestration is next

The second cluster is about getting more than one agent to work together. mvschwarz/openrig, which builds persistent teams out of Claude Code, Codex and Pi, gained more than 2,000 stars this week. The pitch is roles, shared context and owned work: one agent plans, another executes, and they hand off tasks instead of stepping on each other. ruvnet/ruflo takes a similar swing with agent swarms and federated memory. agno-agi/agno offers a platform for building and managing agent systems.

This cluster grew the moment single agents started being useful for real work. One agent is a helper. Two agents that coordinate are closer to a small team, and a small team needs a way to divide labor and keep track of who owns what. That is the gap openrig and its peers are filling.

Skills became a supply chain, and someone shipped the scanner

The third and most telling cluster is skills. mattpocock/skills, a personal collection published as reusable agent skills, pulled in more than 9,000 stars in a week. cursor/plugins publishes an official plugin specification. google/skills and vercel-labs/skills target the same install surface. Several competing repos now compete to be the place you get your skills from.

Whenever an ecosystem grows a distribution channel, a security problem arrives right behind it. That is what makes NVIDIA/SkillSpector the quiet story of the week. It is a scanner that checks an agent skill for vulnerabilities, prompt injection and supply-chain risk before you install it. A scanner only exists when there is something worth stealing or sabotaging, and its appearance is the clearest sign yet that agent skills have become a real supply chain, with all the trust questions that come with one.

The same logic shows up in the auditing tools further down the list. Repos that verify whether an agent actually did what it was asked, and that check its permissions, are climbing for the same reason. Once agents can act, someone has to prove what they did.

Media is the newest entrant

A fourth cluster is smaller but worth noting: agents that make things you can watch. heygen-com/hyperframes, whose one-line description is "Write HTML. Render video. Built for agents," gained more than 4,000 stars this week. calesthio/OpenMontage offers an agentic video production system, and earthtojake/text-to-cad gives an agent the ability to work in CAD. The pattern is that the agent, having learned to write and reason, is now being pointed at pixels and geometry.

Not everything on the list belongs to this story, and that matters for reading it honestly. boykopovar/AnyPS5, a tool for porting PS5 executables to Linux and Windows, pulled in more than 21,000 stars this week and has nothing to do with agents. EpicGames/raddebugger and flutter/flutter are on the list for their own reasons. The AI infrastructure cluster is the largest, not the only, thing happening.

What builders should take from this

Two practical reads follow. The first is that memory has become a buying decision. If you build on agents, you will eventually choose a memory layer, and the options are now mature enough to compare on compression ratio, cross-platform support and how easily you can inspect what got stored. The second is that skills should be treated like dependencies. They carry supply-chain risk, they need versioning, and they deserve the same scrutiny you would give a package from an unfamiliar author.

There is a third read, aimed at anyone deciding what to learn. The skills that are scarce right now are not model training. They are the unglamorous ones: context engineering, permission design, evaluation, and knowing when to give an agent less authority rather than more. The trending page is a rough map of where those skills are needed, and it is more useful for career planning than most job listings, because it shows what people are actually building this week rather than what a recruiter wrote last quarter.

One caveat applies to the whole page. Star velocity is a measure of attention, not quality. Many of these repositories are weeks or months old, and the ones that survive will be the ones that keep shipping after the novelty wears off. The trending list shows where the crowd is looking. It does not promise the crowd is right.

Still, the shape of the list is a useful signal. When the fastest-growing projects are memory, orchestration, skills and auditing rather than flashy demos, it means the agent era has moved past proving it can work and into the unglamorous work of making it reliable. That work is what the trending page is rewarding this week.